摘要
在少数类样本极少的情况下,传统过采样方法难以增加样本数量。为此提出了一种数值生成网络(NGN)算法。该算法将生成器生成的数据作为噪声添加到原始的少数类样本数据中合成新数据,直至达到平衡。生成器采用四层全连接网络,结合低结构和高结构特征生成方法提升数据生成质量和多样性。针对极少数类样本,NGN生成新样本后,与原始少数类数据合并后进行聚类,将聚类后的簇内达到平衡,减小噪声对数据的影响。在6个不平衡数据集上,采用4种算法增加少数类样本使数据集达到平衡,再对所有的数据集使用4种方法分类。实验结果表明,NGN能够增加一定的少数类数量,能够帮助分类模型学习更多少数类的特征,提高分类性能。
When minority class samples are scarce,traditional oversampling methods struggle to increase the sample count.This paper introduced a NGN algorithm that synthesized new data by adding generator-generated data as noise to the original minority class samples until balance was achieved.The generator employed a four-layer fully connected network and integrated low-structure and high-structure feature generation techniques to enhance the quality and diversity of the generated data.For very limited minority class samples,NGN generated new samples,merged them with the original minority class data,and performed clustering to achieve balance within clusters while minimizing the impact of noise.The study evaluated NGN on 6 unbalanced datasets,applied 4 oversampling algorithms to balance the datasets,and classified the balanced datasets using 4 classification methods.The experimental results demonstrate that NGN effectively increases the number of minority class samples,enhances the model’s ability to learn minority class features,and significantly improves classification performance.
作者
邢长征
郑鑫
梁浚锋
Xing Changzheng;Zheng Xin;Liang Junfeng(College of Electronic&Information Engineering,Liaoning Technical University,Huludao Liaoning 125105,China)
出处
《计算机应用研究》
北大核心
2025年第6期1755-1761,共7页
Application Research of Computers
基金
辽宁省教育厅重点资助项目(LJ212410147003)
国家重点研发计划资助项目(2018YFB402900)。
关键词
数值生成网络
生成器
噪声
极少数类
平衡
numerical generator network(NGN)
generator
noise
extremely scarce minority class
balance
